🔍 Read the full analysis: Making An AI Video In One Evening: What I Learned The Hard Way on ThorstenMeyerAI.com
TL;DR
Thorsten Meyer says he built a hand-drawn video engine in one evening, but its first full video was slow and visually weak. After measuring frame use, stillness and camera movement, he added production gates; separate model and monitoring experiments also showed why passing tests or completing tasks do not prove quality.
Thorsten Meyer says he built a hand-drawn video engine in one evening, but the first full video it produced the next morning was, in his words, slow and boring. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how measuring the video’s visual shortcomings led to changes in the engine and checks intended to catch similar problems before later renders.
Meyer said the initial idea came from a hand-drawn explainer he admired. To avoid reproducing it, he says he wrote down a limited set of style traits—hand-drawn strokes, a continuous illustrated world and large numbers used as color accents—and built a separate engine without using the original video’s drawings, assets, layouts or drawing code. The colors came from his own design system, and he said the handwriting font was openly licensed.
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed the code rather than generating it. He said Codex identified five issues in its first review, and each fix was paired with a test designed to fail on the earlier code. Meyer also watched a roughly one-minute test video and approved it for publication; he said that version appeared the same evening in an article about local votes against data centers.
The first full video in the new style was less successful. Meyer said it had too much empty space, icons that did not support the narration and images that stopped moving after they were drawn. He measured the first three scenes before rebuilding the engine. In those scenes, he reported that the share of the frame filled with content rose from 15% to 66%, time spent completely still fell from 29 seconds to zero, and camera moves rose from 3 to 27. Those are figures for the first three scenes, not a stated result for every video.
Infographic 01 · The new video look
One evening, one boring first video, one measured fix
Speed was not the point. The steps in between were: a second AI that reviews, a human who says yes, and numbers that became gates.
Share of the frame filled with content on the first three scenes after the fix. Before: 15%.
From an evening idea to a measured fix
Relative times, this week
- EveningA video I likedClean room: only a list of style traits, no drawings, assets or layouts
- Same eveningOwn engine, first reviewClaude builds; OpenAI’s Codex reviews; 5 findings fixed, each with a test that fails on the old code
- Same eveningA human says yesOne-minute test video approved and published
- Next morningFirst full video: slow and boringToo much white space, icons that did not fit, a picture that stood still
- Hours laterMeasured fix, then gatesThe first video in the improved look goes live
Source: my own build notes for the new video look, this week (engine history, review round, approval, critique, before/after measurement of the first three scenes).
Powered by Thorsten Meyer AI
Turning Visual Critique Into Build Checks
The account highlights a production problem beyond making an AI tool work: a completed render can still fail as a piece of communication. Meyer’s response was to convert subjective criticism into measurements and build gates. He says a scene that is too empty can now be stopped before rendering, while a build can flag an object that does not illustrate what is being said.
That approach may help teams make failures repeatable and easier to catch, but the reported improvements are the author’s own results. The article does not establish that the same thresholds will suit other video formats, audiences or production systems. Its wider lesson is narrower: a model finishing a task, or a test suite passing, is not by itself evidence that the output is useful or correct.
Popular right now
KitchenAid Classic Series 4.5-Quart Tilt-Head Stand Mixer, Onyx Black
Sony BRAVIA 6 65-Inch OLED 4K HDR Smart Google TV with Gemini (K-65A60)
De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Hot and Cold Milk Frother
As an affiliate, we earn on qualifying purchases.
From One Test Render to Published Work
Meyer describes the work as a personal account of taking AI ideas into production, with build notes, incident notes and published experiments as the basis for its figures. He says the short test video was approved and published on the first evening. The weaker full video from the following morning remains online unchanged as an example of what he considers a failed result. A rebuilt video, about the Nasdaq’s worst day in 19 months, went live a few hours after his critique, according to the article.
He connects the video work to a separate experiment involving a locally run model and legal benchmark tasks. Meyer said a weight-edit experiment increased the model’s completion count from 4 to all 10 published tasks, while its legal score remained 0. The account does not provide enough detail in the supplied material to evaluate the benchmark or the scoring method independently; the contrast is presented to show why completion rate alone can mislead.
„“It finishes” is not “it’s right”.“
— Thorsten Meyer
Limits of the Reported Measurements
The figures come from Meyer’s own notes and experiments; the article does not describe an independent evaluation or a comparison against other video-generation systems. It also reports measurements for three scenes without specifying how representative they are of longer videos. The thresholds used by the new gates, and whether they improve audience understanding or viewer response, are not stated in the supplied account.
Meyer also describes a monitoring failure from August: a guard recorded about 1,300 warnings in a log that nobody read, and an alert email did not go out. The account does not specify the exact period covered by those warnings or the cause of the failed email. The supplied text ends while discussing how to judge model quality, so further detail on that experiment is unavailable here.
Gates and Human Review Remain
Meyer says the new frame-coverage and illustration checks are now part of the build process, with a human review still required before publication. The next practical test is whether those gates catch weak scenes consistently in later videos—not just in the examples described—and whether alerts reach a person who can act on them. The article provides no schedule for further testing or independent review.
Key Questions
What did Meyer build in one evening?
He says he built a separate hand-drawn video engine, using a short list of style traits rather than the original video’s artwork or layout.
What was wrong with the first full video?
Meyer described it as slow and boring, citing too much white space, icons that did not support the narration and images that stopped moving after they were drawn.
What changed after the measurements?
For the first three scenes, Meyer reported frame coverage rising from 15% to 66%, stillness falling from 29 seconds to zero, and camera moves increasing from 3 to 27. He says checks based on the measurements became build gates.
Did the separate legal-model experiment show the model was accurate?
Not based on the figures Meyer supplied. The model answered all 10 published tasks, up from 4, but its reported legal score remained 0.
Were the video results independently verified?
The account presents them as Meyer’s own measurements and notes. It does not describe an independent evaluation or provide enough information to compare the results with other systems.
Source: ThorstenMeyerAI.com
